Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Consumer AI assistant customer service results on Zendesk at a glance
- What does the product do?
- Which helpdesk does the team use?
- How did they train their AI customer service agent?
- When did they decide to turn on 'direct replies' to customers?
- What was the biggest thing they did to improve their AI agent's resolution?
- How do they customize their AI agent setup to work for their business?
- Making AI replies feel human
- Guiding tone, context and escalation
- Choosing which tickets the AI replies to
- What impact is the AI customer service agent having now?
- Where do they go from here?
Created time
Jun 3, 2026 09:43 AM
Title length (<60)
Author
Last optimised
Sep 15, 2026
Ecomm?
Image
Publish date
Jun 3, 2026
Video
Slug
consumer-app-ai-customer-support-case-study
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
A consumer AI-assistant app resolves around 40% of roughly 1,200 monthly Zendesk tickets with AI, holding a 77% AI CSAT score and saving about 41 hours every month, with self-learning drafts answering around 1,000 of those tickets in the last 30 days.
When your product lives inside the messaging threads people already use all day, support arrives the same way every other message does: fast, casual, conversational, and constant. For this team, that means a steady stream of repetitive questions all day long, the kind a small team can't keep up with by hand without either hiring or letting replies slip. And because the whole product is built on feeling like a natural conversation, the replies couldn't read like a robot either.
Today our AI agent handles the bulk of that load directly inside Zendesk. It resolves around 40% of roughly 1,200 tickets a month outright, holds a 77% AI CSAT score, and hands the team back about 41 hours a month that used to go on repetitive replies. Around 1,000 of those monthly tickets are answered from help articles the AI drafted for itself.
Here's how it came together.
Consumer AI assistant customer service results on Zendesk at a glance
Fact | A consumer AI-assistant app |
Industry | Consumer software, AI personal assistant delivered through messaging |
Helpdesk | Zendesk (Support inbox, Tickets) |
Support volume | ~1,200 tickets/month |
AI resolution rate | 40% (around 480 tickets a month, capped by their own escalation policy) |
AI CSAT | 77% |
Time saved | ~41 hours/month |
Key features used | Self-Learning, Train on Historic Tickets, Guidance across all three types, AI Tagging, a configured reply delay |
Knowledge sources | Website sync, Notion, uploaded files, historic Zendesk tickets |
Previously evaluated | Zendesk's own AI (ruled out on per-resolution pricing) |
Go-live mode | Direct replies in Zendesk Tickets from the start, with a deliberate reply delay |
What does the product do?
It is a proactive AI personal assistant that works through the messaging apps people already use (think the group chat, where people already are), rather than yet another standalone app to open.
This is an early-stage product with a small team, used by people in their everyday message threads rather than inside a workplace tool.
All of that feeds into support: the questions are conversational, they arrive at consumer volume, and the bar for a reply that feels human is high.
Which helpdesk does the team use?
They run support on Zendesk, in the Support inbox (Tickets), and our approved Zendesk app handles the AI replies inside that same inbox.
They looked at Zendesk's own native AI first and passed on it over price, and I get why. The two products charge in completely different ways.
Zendesk's AI is priced per automated resolution: you're billed each time the AI resolves a ticket on its own. We charge per ticket instead, at around $0.10 a ticket, which works out 3 to 10x cheaper than the per-resolution model.
For a consumer product fielding a high volume of low-value, repetitive messages, per-ticket pricing just made more sense than paying a premium every time a question got resolved.
How did they train their AI customer service agent?
An AI agent is only as good as what it knows, so the team connected several knowledge sources before letting it reply.
Their public website came in through website sync, giving the AI the baseline of what's publicly known about the product. They connected Notion too, so it could draw on the internal product pages the team already maintained.
They also uploaded supporting files with example answers (and here's a wrinkle worth knowing): anything marked private has to be made public before the AI will use it in customer-facing replies. So the team moved the relevant content across.
The highest-leverage source was their own ticket history. With Train on Historic Tickets, we auto-draft knowledge articles from their past Zendesk tickets every week, so the agent learns from how real questions were actually answered, which usually goes well beyond what lives in the help center.
When did they decide to turn on 'direct replies' to customers?
They set the agent up to reply directly to customers inside Zendesk Tickets, but with one deliberate twist. They configured a reply delay, so the AI doesn't fire back an answer the instant a message lands.
What I like about this is the reasoning: an immediate, perfectly-worded reply is a dead giveaway that a bot is on the other end, and for a product built around natural conversation that breaks the spell. A short, human-feeling pause keeps support consistent with the rest of the product.
To let the agent settle in safely, they timed the rollout for quieter, lower-stakes hours, giving Self-Learning room to improve in production before peak volume hit.
What was the biggest thing they did to improve their AI agent's resolution?
The single biggest lever was Self-Learning.
Self-Learning watches what happens after a ticket is handed to a human. It compares the AI's draft to the reply the agent actually sent, and where there's a gap, it drafts a brand-new help article to close it, ready for the team to review.
In the last 30 days alone, those self-written articles fielded close to 1,000 tickets. Paired with the weekly historic-ticket training (the boring-but-effective half of this setup), the knowledge base keeps filling its own holes without anyone sitting down to write articles by hand.
That 40% can look modest next to a number like 1,000, so let me be precise about it (if you want to know what a good rate even looks like, we pulled the numbers from 195 real deployments). The team have chosen where to draw the line, and the AI isn't bumping into a ceiling.
We count a ticket as resolved when it's handled without being escalated to a human, and this team deliberately escalate a lot. Anything sensitive or personal gets routed straight to a person by their own handover guidance.
So Self-Learning does the heavy lifting on the volume of questions answered, while the team keep a tight, self-chosen policy on what the AI is allowed to close on its own. The 40% is a line they've chosen to draw; the AI didn't run out of road.
How do they customize their AI agent setup to work for their business?
A consumer product with a distinct personality can't run a generic bot, so the team shaped the agent around a few things in particular.
Making AI replies feel human
The reply delay is the clearest example. The agent waits a beat before answering, so messaging support gives a customer the same unhurried feel they get from the product itself.
It's a small setting that does a lot of work for a brand whose whole pitch is that talking to it feels natural.
Guiding tone, context and escalation
They use Guidance across all three types. Communication and Style guidance keeps the tone on-brand. Context and Clarification guidance tells the agent when to ask for more detail before answering.
Handover and Escalation guidance defines exactly when the AI should step back and pass a ticket to a person. That rule sets the 40% figure. Escalations are tracked with a handover tag in Zendesk and show up in our Inspect view, so the team can see precisely what the AI is passing on and why.
Choosing which tickets the AI replies to
They also use AI Tagging in Zendesk to auto-classify incoming tickets and keep the AI away from categories it shouldn't touch. Sensitive topics like legal queries get tagged and blocked from AI replies (yes, even when the AI could probably answer), so those conversations go straight to a human.
It's a clean way to draw the boundary of what the agent is allowed to handle.
What impact is the AI customer service agent having now?
The numbers from the last 30 days:
- ~40% AI resolution rate: a deliberate line the team set with their own escalation guidance.
- ~1,200 tickets handled per month, of which around 480 are resolved by the AI without a human.
- ~41 hours saved per month: at roughly five minutes of agent time per ticket, those resolved tickets add up to about 41 hours the team didn't have to spend.
- 77% AI CSAT across the tickets the AI handled.
- ~1,000 tickets a month answered from help articles Self-Learning drafted automatically.
The tickets the AI doesn't auto-resolve are the ones the team's guidance deliberately sends to a person. That's why I'd trust this 40% more than a flashier one.
Where do they go from here?
The clear next step is live customer data. They haven't yet connected the User Data API, which would let the AI answer account-specific questions, like what's happening on a particular user's account, from real data.
For a consumer product, that's the unlock that turns a lot of today's escalations into things the AI can handle on its own. From there, Tasks and Tools would let the agent move from answering to doing: taking actions on a user's account through the product's own APIs, with the right approvals in place.
'Fully automated' is easy to oversell for a product like this. For them, it means the AI owns the repetitive, conversational core, the questions that arrive in the same form all day, while the team keeps the judgment calls and anything personal or sensitive.
That split is why a deliberately controlled 40% comes with a 77% satisfaction score.
This setup echoes what we've seen with other teams: Honeygain on the same Zendesk stack, and a high-volume consumer platform facing the same kind of repetitive consumer volume. If you'd like to see more stories like this one, browse all our case studies, or check our pricing to model what this would cost at your ticket volume, or run your own numbers through the Zendesk ROI calculator.
